When dist test was updated to run "go test" with multiple package
arguments at once, merging the logical test units into one execution,
the hack to give cmd/go twice as much time wasn't updated.
What was happening (even in the all.bash case) was that we were
merging together, say, "cmd/go" and "bytes", and because bar was
lexically earlier, the timeout calculation was based on package "byte",
even though we were actually running, say: "go test bytes cmd/go".
This explains why x/build/cmd/release was often flaky with its
all.bash, since cmd/go can't really finish in 3 minutes reliably
unless it's running by itself. If it has any competition, it runs
over.